Cbinsights-2022年人工智能趨勢值得關(guān)注 AI trends to watch in 2022

2022

AItrendsto

watchin2022

Whereisallthisdatafrom?

TheCBInsightsplatform

hastheunderlyingdata

includedinthisreport

Clickheretosignupforfree

TableofContents

Syntheticdataforprivacy.Fromhealthcaretofinance,

corporationsareexperimentingwithsyntheticdatasetsto

enabledatasharingandcollaboration—whilecomplying

withGDPRandotherprivacylaws.

The$67Bchiprace.Ascompaniesscrambletodominate

inAIchips,newinnovationsfrom3Dchipstoin-memory

computingarepoisedtoshakeupthespace.

Protectingvirtualworlds.Thethreatoftoxiccontentand

behaviorhasspilledoverfromsocialmediatoanewfront:

onlineworlds.CompaniesareusingAItodetectharmful

behavioringamingandothervirtualspaces.

Weedingoutdeepfakes.Asdeepfakecandidates

campaignforvotesandAIvideosspreadwartime

misinformation,deepfakedetectiontechcompeteswith

theAIcreatingit.

Augmentedcoding.AIismakingcodingeasierandmore

efficientbyassistingwithcodecompletionandautomating

softwaretesting

.

MultimodalAI.Techcompaniesaremakingstridesin

multimodalAI—asingleAImodelthatcanunderstand

conceptsfrommultiplemodalities,likevideo,text,and2D

images—toimprovesearchandcontentgeneration.

End-to-endmachinelearning.Inthefragmentedmarket

forAIdevelopment,end-to-endmachinelearningplatforms

thattakeaprojectfromstarttofinishwillemergeas

acompellingoptionforenterprisesstrugglingwithAI

implementation

.

6

9

13

16

20

22

25

AITRENDSTOWATCHIN20223

Intro

Theeraofmachinelearninghasarrived.

AIcompaniesraisedarecord

$66.8Binfunding

in2021—over

doubletheglobalfundingtotalin2020.AndwhilequarterlyAI

funding

slipped12%fromQ4’21toQ1’22

,thisdeclinewasless

dramaticthanventureasawhole,whichsawfundingfall19%

duringthesameperiod.

Inour

2021enterpriseAItrendsreport

,wehighlightedthemesthat

continuetoremainrelevant:naturallanguageprocessingmodels

aregettingbiggerandbetter,AIOpsvendorsareraisingmulti-

million-dollarfundingrounds,andmorevendorsareintegrating

no-codefeaturesintotheirplatforms.

Now,newtrends—aswellasnewsolutionstooldproblems—that

arerelevanttoconsumersandenterprisesalikeareappearing

acrossindustries.

The“metaverse”hasemergedasamajorthemeinthelastyear,as

onlineworldsandgamesgrewinpopularityandFacebookmadeits

highlypublicizednamechangetoMeta.Thishasinturnfueledthe

needfornewAI-poweredcontentmoderationtoolstomonitorfor

toxicandhatefulbehavioronline.

FromthewarinUkrainetopresidentialcampaignsinSouthKorea,

deepfakeshavebecomemoreprominentinmainstreammediaand

politics—puttingdeepfakedetectionandmediaprovenancetech

backinthespotlight.

AITRENDSTOWATCHIN20224

Finally,tocomplywithdataprivacylaws,corporationsacross

healthcare,telecommunications,andfinanceareturningto

syntheticdata:fakedatathatcloselymimicsreal-worlddata.

Fromfamiliartechnologytocutting-edgeadvancements,readon

for7AItrendstowatchcloselythisyear.

AITRENDSTOWATCHIN20225

Syntheticdataforprivacy

Fromhealthcaretofinance,corporationsareexperimentingwith

syntheticdatasetstoenabledatasharingandcollaboration—

whilecomplyingwithGDPRandotherprivacylaws.

Inindustrieswheresufficientreal-worlddatatotrainAIis

unavailableorsparse,orwherecomplianceandprivacyaremajor

concerns,enterprisesareturningtosyntheticdatasets:fake

images,videos,ortabulardatathatresemblereal-worlddatasets.

Despitethechallengesinaccuratelymimickingreal-worlddata,big

corporationsarebettingonthetech.

SOURCE:

CBINSIGHTSESPVENDORMATRIXFORSYNTHETICTABULARDATAPROVIDERS

Illuminaisusingsyntheticgenomicsdatadevelopedbystartup

Gretel

formedicalresearch.Inajointcasestudy,the2companies

highlightedissueslikerestrictivelegislationandpatientconsent

requirementsthatlimitthespeedandscaleofmedicalresearch

thatreliesonsensitivepatientdata.Inresponse,Gretelusedreal

genotypeandphenotypedatatotrainanAIalgorithmtogenerate

artificialgenomicdata.

AITRENDSTOWATCHIN20226

“Alongwithprivacyenhancingtechnologiessuchas

differentialprivacy,syntheticdatahasthepotential

toaddressthedeepprivacyconcernsworkingwith

genomicdata,enablingfastersharingofdataand

unlockinginnovation.”

—Gretel-Illuminajointcasestudy

Inthefinancialsector,J.P.MorganistrainingfinancialAImodels

withfakedata,whileinthetelecomindustry,Telefónicahasworked

with

MostlyAI

tocreateGDPR-compliantsyntheticcustomer

profilesthatmirrorstatisticalpatternsofrealcustomerdata.It’s

estimatedthatupto85%ofrealcustomerdatainthetelecom

industryisunusableduetolackofcustomerconsent,making

behavioranalysisandpredictiondifficult.

InMarch2022,

MDClone

raised$63MfromLightspeedVenture

Partners,OrbiMedAdvisors,andothersforitssynthetichealth

dataplatform.Thecompanyclaimsthatsyntheticdataisaviable

alternativetoothermethodsofdatade-identification(i.e.,removing

personallyidentifiableinformationfrompatientrecords)where

patientsmaypotentiallybere-identifiedthroughcross-referencing.

Anothervendorinthespace,

Tonic

,hasintroducedAIfeatureson

itsplatformtoimprovethefidelityoffakedata.Itraised$35Mlast

yearfromInsightPartners,BloombergBeta,andothers,andworks

withclientsincludingFlexport,eBay,andOscar.

AITRENDSTOWATCHIN20227

Whilesyntheticimageandvideogenerationhastakenoff—thanks

to

GANs

,anowpopularapproachthatuses2neuralnetworksto

generatehyper-realisticsyntheticvideosandimages—synthetic

tabulardatagenerationisstillinitsearlystages.Buttheupsidesof

dataanonymization,privacycompliance,andpotentialbiasreduction

makeitanappealingoptionforcorporationsacrossindustries.

AITRENDSTOWATCHIN20228

The$67Bchiprace

AscompaniesscrambletodominateinAIchips,newinnovations

from3Dchipstoin-memorycomputingarepoisedtoshakeup

thespace.

TherapidcommercialsuccessofAIacrossindustriesiscreating

demandforspecializedhardwarethatcansupportcompute-

intensiveAIworkflows,bothinclouddatacentersandedgedevices

likecameras.

OncedominatedbyNvidia’sGPUs,the

$67BAIchipmarket

has

beenblownopenbyanumberofplayers.

Google’slatestPixelphonesarepoweredbythetechgiant’sown

Tensorprocessorstosupporton-deviceAIapplications,while

AmazonlaunchedGraviton3,itscustomchipforAIinference,in

Q4’21.

Beyondtheincumbents,startupsarealsoenteringthefray.

CerebrasSystems

offerswhatitclaimstobe“thelargestchipever

built,”with2.6Ttransistorsand850,000AIcores.Thecompany

raised$250MinQ4’21ata$4Bvaluation.

AITRENDSTOWATCHIN20229

However,mammothchipsaren’tpracticalformanyeverydayAI

applications,duetospaceandenergylimitations.Seeingthis

opportunity,agrowingnumberofcompaniesareofferingAIchips

thatcanbeusedwithlow-powerdeviceslikecarsensors,store

cameras,automatedfactoryrobots,andmore.

On-deviceAIprocessingcanpowerlow-latency—andpotentially

privacy-preserving—experiences.Startupsincluding

Mythic

,

Syntiant

,and

Kneron

haveeachraisedmorethan$100Mtodevelop

thistechnology.

SOURCE:SYNTIANT

Meanwhile,startupslike

UntetherAI

and

HOUMO.AI

areworking

on“in-memorycomputing.”ThisapproachbringsAIprocessors

andmemorytogetherontoasinglechip.Thecloseproximityand

highlevelofintegrationinthesesystemscandeliverdramatic

performancegainscomparedtoconventionalmethods:Samsung

reportsitusedtheapproachtomorethandoublethespeedofa

speechrecognitionneuralnetwhilecuttingitsenergyconsumption

inhalf.

AITRENDSTOWATCHIN202210

SOURCE:GRAPHCORE

AIchipunicorn

Graphcore

isalsoplayingwithchipstructureto

boostperformancegainsusinganapproachcalled“3Dchips.”

Thetechinvolvesbondingmultiplechipsetstogethertocreatea

fusedstackofintricatelyalignedcomponents.Graphcoreused

thismethodtoplaceapowerdeliverychipdirectlyontopofanAI

processor,allowingthechiptotrainaneuralnet40%fasterthanits

previousgenerationwhileusinglessenergy.

OthercompaniesareditchingthephysicsofconventionalAI

chipsaltogetherbyturningtophotonicprocessors,whichshuttle

dataaroundusinglightasopposedtoelectricalsignals.Thebig

advantageofphotonicsisspeed—lightcanmoveinformation

faster,withgreaterbandwidth,andusinglessenergythanrelying

purelyonelectronsflowingthroughcables.Withthedemandsof

AItoolsexpectedtocontinuegrowingexponentially,photonics

promisesawaytoavoidhittingaloominghardwarebarrier.

AITRENDSTOWATCHIN202211

Researchershavestruggledfordecadestomakepractical,general-

purposechipsusingphotonics.However,startupsincluding

Lightmatter

and

LuminousComputing

arebettingthattheycan

usephotonicchipsoptimizedfortaskslikedeeplearningtobuild

“AIsupercomputers”capableofhandlingevenparticularlytaxing

algorithms.

SpecializedAIchipsareheretostay—andincumbentslikeNvidia

andIntelarefacingfiercecompetitionfromstartupsandbigtech

leadersalike.Newtechnologiesareemergingthatwillmakethese

chipsperformbetteranduselessenergy.

Lookingfurtherahead,cutting-edgetechnologylikequantum

machinelearning—whichusesquantumcomputerstohelprun

certainAItasksmuchmoreefficiently—willcomplicatethis

landscapeevenfurther.

AITRENDSTOWATCHIN202212

Protectingvirtualworlds

Thethreatoftoxiccontentandbehaviorhasspilledoverfrom

socialmediatoanewfront:onlineworlds.CompaniesareusingAI

todetectharmfulbehavioringamingandothervirtualspaces.

IntheUS,thenumberofpeopleplayingvideogamesisatanall-

timehigh.Additionally,76%ofAmericansunder18playgames,

raisingconcernsthatchildrenwillbeexposedtoinappropriateor

hatefulcontent.

Detectinghatespeechonlineisnotanewproblem.Metareportedly

spent$13Boncontentmoderationbetween2016and2021,while

TikTokclaimedin2020tohave10,000peopleworkingoncontent

moderation.BothTikTokandMetahavebeensuedbycontent

moderatorswhosufferedpsychologicaldamageonthejob.

Now,giventheincreasedhypearoundthemetaverseandarapidly

evolvingonlinegamingecosystem,thethreatoftoxiccontentand

behaviorhasspilledoverfromsocialmediasitestoanewfront:

onlinegamingandvirtualworlds.

AITRENDSTOWATCHIN202213

Gamescanquicklyturntoxicintheformofhatespeech,

cyberbullying,griefing(intentionallyhelpingtheotherteamwin

tofrustrateyourteammates),andragequitting(refusingtoplay

agametoruintheexperienceforeveryoneelse).Astudyfrom

theAnti-DefamationLeaguefoundthatupto80%ofplayers

haveexperiencedharassmentinsomeofthemostpopular

multiplayergames.

Startupsarebothpivotingandemergingtoaddressthismarket

withthebeliefthatAIcancombatthisbehavioratscale.

Spectrum

Labs

claimsitsnaturallanguageprocessing(NLP)platform

reducesaudio-andtext-basedcontentmoderationeffortsby50%

whileimprovingtoxicbehaviordetectionby10x.

Meanwhile,inMarch2022,

GGWP

emergedfromstealthand

announceditsintentiontocombattoxicbehavioringamesby

usingAItomonitorchatlogsandgameplaydata.Thestartup

hadpreviouslyraiseda$12Mseedroundbackedbyinvestorslike

BITKRAFTEsportsVentures,SonyInnovationFund,andRiotGames.

InApril2021,

Hive

raiseda$50MSeriesData$2Bvaluation.The

startup,whichwasfoundedin2013asaplatformtocrowdsource

AItrainingdata,isprovidingautomatedcontentmoderationtohelp

metaversecompanieslike

HighRise

and

NaverZ

(whichrunsthe

Zepetoplatform)stayontopofplayerhatespeech,cyberbullying,

andmore.

AITRENDSTOWATCHIN202214

SOURCE:HIVE

BigtechplayersarealsosnappingupAIstartupstobring

moderationcapabilitiesinhouse.InOctober2021,forinstance,

Microsoftacquired

TwoHat

,anAIcontentmoderationservicethat

listsRoblox,EpicGames,andMicrosoft-ownedMinecraftasits

customers.

Ultimately,perfectcontentmoderationmaybeimpossible.Online

communitiesconstantlyadapttoescapeplatformcensorship.

However,breakthroughsinkeyareaslikeNLPandimage-based

classification,aswellasrecentfundingroundstoAI-powered

contentmoderationstartups,indicatethatAIwillbeattheforefront

ofthecontentmoderationwar.

AITRENDSTOWATCHIN202215

Weedingoutdeepfakes

AsdeepfakecandidatescampaignforvotesandAIvideosspread

wartimemisinformation,deepfakedetectiontechcompeteswith

theAIcreatingit.

Deepfakes

haveexpandedfromhyper-realisticimagestovoiceand

videospoofs—includingfacereenactment,whereonefaceina

videoisswappedoutwithanother.

Poweredbyself-learningalgorithms,thesefakeshavegottenbetter

andbetterovertime.Thesheernumberofpubliclyavailablevideos

andvoicerecordingsmakesiteasytotrainAIalgorithmsand

generatedeepfakes—especiallyofcelebrities.Researcherssayit

willonlygethardertodistinguishAI-generatedfaces,objects,and

videosfromrealones.

Deepfakeshaveproliferatedinthemediain2022,especiallyinthe

politicalsphere.InMarch2022,TheWallStreetJournalreported

howYoonSuk-yeol,acandidateintheSouthKoreanpresidential

elections,wasusingdeepfakesofhimselftoimprovehispublic

imagewithyoungvoters.The“AIYoon,”asthedeepfakeversionwas

called,cameoffasmorehumorousthanitsreal-lifecounterpart.

YOONSUK-YEOL.SOURCE:THEWALLSTREETJOURNAL

AITRENDSTOWATCHIN202216

InEurope,viraldeepfakesofVolodymyrZelenskyyandVladimir

PutinhaveaddedtowartimemisinformationinUkraineand

Russiainrecentmonths.Andin2021,awaveofpoliticiansaround

EuropeweretrickedduringvideocallsbylivedeepfakesofRussian

politicalactors.

Beyondfakenewsandpoliticaldisinformation,deepfakesarealso

targetingenterprisecustomers—andhavethepotentialtobecome

amainstreamtoolinconsumerphishingandextortionscams.

“Whatweobserveisthatfakemediaisgettingbetter

andbetterandbetter,andtheabilityforcomputersto

tellwhat’srealandwhat’sfake,oruserstotellwhat’s

realandwhat’sfake,israpidlytendingtozero.”

—PaulEngland,engineeratMicrosoftResearchLab

ArguingthatAI-powereddefensecannotfullycombatAI-generated

fakes,Microsoft(alongsidepartnerslikeAdobeandtheBBC)

launchedProjectOrigin,aservicethatallowspublishersto

authenticatemediausingtamper-proofmetadata.

Microsofthasalsoseparatelypartneredwithstartup

AIFoundation

,

whichoperatesbothcommercialandnon-profitventures.AI

Foundation’snonprofitisheadedbyTwitterco-founderBizStone,

TrionWorldscreatorLarsButtler,andothers.

AITRENDSTOWATCHIN202217

In2020,AIFoundationraised$17Mtobuilddeepfakeavatars,

whilealsoannouncingthelaunchofadeepfakedetectionplatform,

RealityDefender,withMicrosoft.RealityDefenderhaspartnerships

withABC,HomelandSecurity,andtheDepartmentofDefense,

amongothers.

Takingadifferentapproach,

Truepic

usescryptographyand

blockchain-basedtechforphotoandvideoidentification.Truepic

hasbeenworkingwithQualcommtoaddacryptographic

sealaroundpicturescapturedbyQualcommchipset-powered

smartphones.Itraised$27MlastyearfromAdobe,Microsoft’s

M12,SonyInnovationFund,andothers.

SOURCE:METAAI

Lastyear,researchersatMetaannouncedthattheycannow“reverse

engineer”adeepfakeimage—thatis,notjustdeterminewhetheran

imageisfakeornot,butalsodissecttheattributesoftheAImodel

usedtocreatethedeepfake.Thismakesitpossibletodetermine

whethermultiplefakemediaoriginatedfromasinglesource.

AITRENDSTOWATCHIN202218

Asafirstwaveofresponsetothisgrowingcybersecuritythreat,

techcompanieshaveexperimentedwithapproachesrangingfrom

on-deviceauthenticationsoftwareandAPIstoblockchain-based

solutions.However,deepfakeswillcontinuetobecomemore

advancedandubiquitous,forcingtechcompaniestofindnewways

todetectanddisruptthem.

AITRENDSTOWATCHIN202219

Augmentedcoding

AIismakingcodingeasierandmoreefficientbyassistingwith

codecompletionandautomatingsoftwaretesting.

Algorithmsaretakingnaturallanguagecommandsandtranslating

themintocomputercode,signalinganeweraforsoftware

development—especiallyamongcitizendevelopers.

InJune2021,

GitHub

(acquiredbyMicrosoftin2018)and

OpenAI

(inwhichMicrosoftownsa$1Bminoritystake)tag-teamedto

launchGitHubCopilot.TrainedonpubliclyavailableGitHubdata,

CopilotconvertsEnglish-languagecommentsintocode.Copilot

isparticularlyusefulinmappingsmallerprogrammingtasksto

existingcode.

SOURCE:GITHUBCOPILOT

AITRENDSTOWATCHIN202220

Microsoftisn’ttheonlybigtechplayermakingmovesinthespace.

InFebruary2022,Google’s

DeepMind

releasedAlphaCode,AI-

generatedprogramsthatweretestedonmorecomplexproblems

incompetitiveprogramming.AfterevaluatingitsAIonrecent

competitionshostedonCodeforces,DeepMindclaimedthatitsAI

performs“roughlyatthelevelofthemediancompetitor.”

Ontheprivatemarketside,startupsaremorefocusedonsoftware

testing:automatingqualitychecksandunittestsofcode.Key

vendorsinthespaceinclude

Mabl

(backedbyCRVandGoogle

Ventures),

Autify

(backedbySalesforceVentures),andUniversityof

Oxfordspinout

Diffblue

(whichraiseda$10MSeriesBinQ1’22).

Automatedprogrammingisstillinitsinfancy.Butrecentadvances

favoragrowingenterprisetrendtowardno-code/low-code

solutionsthatenablenon-technicaluserstoparticipateindata

scienceprojects,closeskillsgaps,andspeedupproductioncycles.

AITRENDSTOWATCHIN202221

MultimodalAI

TechcompaniesaremakingstridesinmultimodalAI—asingle

AImodelthatcanunderstandconceptsfrommultiplemodalities,

likevideo,text,and2Dimages—toimprovesearchandcontent

generation.

Today,anAImodeltrainedonvideodatacanbeusedforpredicting

videocontent,amodeltrainedontextcanbeusedfortext

predictions,andsoon.

Togobeyondthis,multimodalAIresearchaimstobemoreholistic,

usingasingleAImodeltoconceptualizeinformationacrossmultiple

typesofdataliketext,2Dimages,andvideostomakeaprediction.

Forexample,inearly2021,OpenAItrainedanAImodelcalled

DALL-Etogenerateimagesbasedonatextprompt.Intheimage

below,theAIgeneratesavocado-shapedarmchairsfollowinga

promptforthesame.

SOURCE:OPENAI

InJanuary2022,OpenAIreleasedDALLE-2,whichimprovesthe

originalmodel’soutputimageresolutionby4x.

InMay2022,GooglelaunchedImagen,atext-to-imageprojectthat

reportedlyoutperformsOpenAI’smodelintermsofthequalityof

imagesgenerated,aswellasthealignmentbetweentheinput(text)

andoutput(AI-generatedimage).

AITRENDSTOWATCHIN202222

SOURCE:GOOGLE

Earlierthisyear,Metapublishedapapercalled“Omnivore:ASingle

ModelforManyVisualModalities.”ThepaperdetailsanAImodel

that,whentrainedtorecognize2Dimagesofpumpkins,canalso

recognizepumpkinsinvideosor3Dimageswithoutrequiring

additionaltrainingforthelatter2mediatypes.

MultimodalAIisgrowingbeyondacademicresearchlabstofind

practicalapplications.

Google,forinstance,isusingmultimodalAItoimprovesearch.In

thefuture,ausercouldtakeaphotooftheirhikingbootsandask

aquerylike,“CanIusethesetohikeMt.Fuji?”Thesearchengine

wouldrecognizetheimage,mineinformationonthewebaboutMt.

Fujifromtext,image,andvideodata,andconnectthedotstoprovide

arelevantanswer.

AITRENDSTOWATCHIN202223

SOURCE:GOOGLE

Onthestartupside,

TwelveLabs

raised$5Minseedfundingfrom

IndexVentures,AIresearcherFei-FeiLi,andotherinvestorsin

March2022.ThecompanyisdevelopingAItounderstandcontext

frombothvisualsandaudiodatatoindexvideosandmakethem

searchable.

MultimodalAIresearchispoisedtogobeyondcorporateresearch

labstopowerthenexteraofsearchandcontentgeneration,among

otherapplications.

AITRENDSTOWATCHIN202224

End-to-endmachinelearning

InthefragmentedmarketforAIdevelopment,end-to-endmachine

learningplatformsthattakeaprojectfromstarttofinishwill

emergeasacompellingoptionforenterprisesstrugglingwithAI

implementation.

AscommercialapplicationsofAIscalerapidly,enterprisesare

lookingtooverhaulexistingdatamanagementpracticestobecome

“AI-f